3.1. The Power of Gap Characters to Resolve the Avian Tree of Life
Resolving the topology deep in the avian tree of life is a notoriously difficult problem [
12,
76,
77], making it an excellent test case for novel sources of phylogenetic information. The power of specific types of data to resolve phylogenetic relationships depends upon the size of the matrix, rate of evolution, amount of homoplasy and branch lengths in the true tree. The ideal evolutionary rate for phylogenetic characters is rapid enough for a high probability of synapomorphic changes to occur on the shortest branches in the tree, but not so high that homoplastic changes obscure historical signal [
36,
37]. The nucleotide substitution rate for introns appears to be appropriate for analyses of deep avian phylogeny [
12]. In contrast, gap characters accumulate at a much lower rate (the MP treelength given gap data are approximately 10% of the treelength given nucleotide data). The ML estimate of the gap accumulation rate is even lower (
Figure 2), although the lower homoplasy of gap characters may prove advantageous if very large gap datasets were analyzed.
Figure 2.
Branch lengths estimated from gap data (using the CFNv+Γ model) plotted against branch lengths from all nucleotide data (estimated using the GTR+I+Γ model). Branch length estimates for specific nucleotide partitions (introns, coding exons and 3' untranslated regions [UTRs]) are presented for comparison of relative rates (next page).
Figure 2.
Branch lengths estimated from gap data (using the CFNv+Γ model) plotted against branch lengths from all nucleotide data (estimated using the GTR+I+Γ model). Branch length estimates for specific nucleotide partitions (introns, coding exons and 3' untranslated regions [UTRs]) are presented for comparison of relative rates (next page).
To examine the phylogenetic signal in gap characters, we obtained estimates of the avian tree of life based only upon gap characters (
Figure 3 and supporting information, files 3 and 4). The gap tree had relatively high bootstrap support for most orders (
Figure 3), the structure within orders (supporting information, files 3 and 4) and the small number of strongly supported supra-ordinal clades (
i.e., the clades indicated with red asterisks in
Figure 1), albeit often with lower bootstrap support than the nucleotide tree. Those supra-ordinal groups recovered in the gap trees (e.g., Novaeratitae, Picocoraciae, Picodynastornithes and Strisores) were much more poorly supported by the bootstrap in the gap character tree than they were in the nucleotide tree. Other independently corroborated supra-ordinal clades were not even present in the gap tree (e.g., Telluraves). However, there was also an interesting exception; McCormack
et al. [
64] found a strongly supported Eurypygiformes-Phaethontiformes clade. This clade is present in the gap trees. We have refrained from suggesting a name for this clade, since it is absent from the Early Bird tree and lacks independent corroboration, but it could be a case where analyses of gap characters exhibit better agreement with other sources of information than the analyses nucleotides conducted by Hackett
et al. [
13]. Overall, these analyses demonstrated that a large gap character matrix has sufficient phylogenetic signal to recover many of the most strongly corroborated nodes in the avian tree of life, but few of the most difficult nodes.
Substantial branch length heterogeneity was evident in both the nucleotide and gap trees, and branch lengths appear to be somewhat correlated between the two data types (
Figure 4). Several taxa have long branches relative to their close relatives, including
Turnix (Charadriiformes), Tinamiformes (Paleognathae) and Phasianidae (represented here by the genera
Coturnix,
Gallus and
Rollulus within the order Galliformes), in both the nucleotide and gap trees (
Figure 4). This indicates that rates of nucleotide substitution and the accumulation of gap characters are correlated in birds, as expected based upon analyses of other groups of organisms (e.g., Hardison
et al. [
78]).
This branch length heterogeneity may influence the estimate of topology, and it is tempting to speculate that the clustering of the long-branched Psittacopasserae and Picocoraciae within Telluraves reflects long branch attraction, especially given the short branches associated with the raptorial taxa (Accipitriformes, Cariamiformes, Falconiformes and Strigiformes) within this supra-ordinal clade. If so, the gap tree would actually provide a less accurate estimate of avian phylogeny than the nucleotide tree given that both Psittacopasserae and Australaves are paraphyletic in the ML gap tree but strongly supported by independent evidence [
30,
60].
The observed branch length heterogeneity suggests that ML methods might provide better estimates of avian phylogeny than MP, because parsimony equivalent models (
i.e., the “no common mechanism” [NCM] model [
79]) are unlikely to account effectively for branch length heterogeneity [
80,
81]. Indeed, it is clear that standard model selection approaches will indicate that the CFNv+Γ model has a better fit to the data than NCM [
80], although we do note that there is debate regarding the question of whether MP should be viewed as a model [
82]. Despite this prediction, our results are equivocal regarding the relative performance of these methods (e.g., compare
Figure 3A to 3B). Indeed, the MP tree supports monophyly of Psittacopasserae and Austrodyptornithes (
Figure 3B), unlike the ML tree, albeit with low (<50%) bootstrap support in both cases. The best interpretation of these differences between the MP and ML topologies is unclear, although differences between the trees at the supra-ordinal level provide no clear evidence that ML using the CFNv+Γ model performs substantially better than MP.
Figure 3.
Estimates of avian phylogeny obtained using 12,030 gap characters obtained using (
a) ML analyses with the CFNv+Γ model and (
b) the maximum parsimony (MP) criterion. Orders were collapsed when monophyletic to simplify the trees. Bootstrap support on terminal branches reflects the support of those orders; orders represented by a single taxon are indicated using “(1)”. There were a limited number of rearrangements relative to the nucleotide topology within orders, most without bootstrap support. We highlighted the topology for the order Galliformes, because the gap topology included a clade with bootstrap support that conflicts with multiple nuclear gene regions [
8,
83] and morphology [
84].
Figure 3.
Estimates of avian phylogeny obtained using 12,030 gap characters obtained using (
a) ML analyses with the CFNv+Γ model and (
b) the maximum parsimony (MP) criterion. Orders were collapsed when monophyletic to simplify the trees. Bootstrap support on terminal branches reflects the support of those orders; orders represented by a single taxon are indicated using “(1)”. There were a limited number of rearrangements relative to the nucleotide topology within orders, most without bootstrap support. We highlighted the topology for the order Galliformes, because the gap topology included a clade with bootstrap support that conflicts with multiple nuclear gene regions [
8,
83] and morphology [
84].
Figure 4.
Branch length heterogeneity evident in the (a) optimal nucleotide tree (based upon the GTR+I+Γ model) and (b) the optimal gap tree (based upon the CFNv+Γ model).
Figure 4.
Branch length heterogeneity evident in the (a) optimal nucleotide tree (based upon the GTR+I+Γ model) and (b) the optimal gap tree (based upon the CFNv+Γ model).
Examining other aspects of model fit, including conducting ML analyses without correcting for acquisition bias (i.e., using the CFN+Γ model), also resulted in similar topologies. These equivocal results are most likely to reflect the limited phylogenetic information in gap data matrices, even ones as large as that analyzed here. This suggests that it will be necessary to examine even larger data matrices to determine whether either analytical approach provides an adequate fit to the underlying process of indel evolution and to establish the impact of these methods upon topology.
3.2. Phylogenetic Signal in Gap Characters Based upon Indels of Different Lengths
Above, we described two reasons why gap trees might have lower bootstrap support than the nucleotide tree. Specifically, the limited bootstrap support we observed could reflect the low rate of accumulation for gap character changes or poor model fit (alternatively, it could reflect a combination of both). Another possibility is that the gap data are sufficiently noisy that neither ML nor MP can recover an accurate estimate of the true tree. Even if noise is not positively misleading, it can have a negative impact upon the phylogenetic analyses [
85]. Thus, noise reduction methods might provide a useful complement to model improvement. Short indels, especially 1-bp indels, are more common than long indels in avian non-coding regions [
10,
86], suggesting that gap characters based upon short indels may contain more noise than those based upon long indels. Thus, the removal of short indels has the potential to enhance phylogeny reconstruction.
To examine the utility of noise reduction based on gap length, we filtered the full gap data matrix (12,030 characters of which 4,245 were parsimony informative) and excluded gap characters based on short (1- and 2-bp) indels. Removing 1-bp gaps reduced the matrix size by almost 25% (to 9,115 characters; 3,160 parsimony informative), whereas excluding both 1- and 2-bp gap characters reduced the matrix size by an additional 11% relative to the original matrix size (to 7,740 characters; 2,640 parsimony informative). Although the rate of longer gap accumulation was lower (the rate after excluding 1-bp gaps is 76% of that for the all gap matrix and the rate after excluding 1- and 2-bp gaps is 64%) all three data matrices exhibit similar levels of homoplasy (
Table 2). Estimates of phylogeny obtained after removing short indels did not improve congruence with the nucleotide data tree (
supplementary information, file 2). Robinson-Foulds distances [
87] between the nucleotide trees and all of the gap trees ranged from 92 to 100, whereas the distance among gap trees ranged from 64 to 70. Removing short gaps may prove beneficial for other data sets, but these results showed that 1- and 2-bp gaps did not contribute substantially to the noise in the gap dataset.
Table 2.
Retention indices [
88] for gap characters and nucleotide data. Retention indices were calculated using the ML topologies for nucleotides (
Figure 1) or gaps (
Figure 3a).
Table 2.
Retention indices [88] for gap characters and nucleotide data. Retention indices were calculated using the ML topologies for nucleotides (Figure 1) or gaps (Figure 3a).
| Topology |
---|
Data Matrix | Nucleotide tree | Gap tree |
Gaps | | |
| All | 0.7154 | 0.7209 |
| >1-bp (excluding 1-bp gaps) | 0.7141 | 0.7190 |
| >2-bp (excluding 1- and 2-bp gaps) | 0.7238 | 0.7288 |
Nucleotides | | |
| All | 0.5231 | 0.5188 |
| Introns | 0.5206 | 0.5167 |
| Coding exons | 0.5315 | 0.5251 |
| 3' untranslated regions | 0.5632 | 0.5597 |
Not surprisingly, given the similar level of homoplasy in the full gap-data matrix and the filtered matrix with 1-bp gaps removed, bootstrap support in analyses using identical numbers of informative characters was similar in trees made from both data sets (
Figure 5a). In fact, only four nodes exhibited fairly large changes in bootstrap support when 1-bp gaps were removed. In three cases, this was an improvement (from <50% to ≥70%); in the fourth case it was a decrease (from 69% to 12%). The node with reduced support united Picodynastornithes, a clade with independent corroboration (
Table 1). In fact, Picodynastornithes was not present in the ML tree for gap data excluding 1-bp gaps; instead, the ML tree included a conflicting clade that comprised Coraciiformes and Bucerotiformes (supporting information, files 3 and 4). Similar results were obtained when both 1-bp and 2-bp gaps were excluded.
Surprisingly, the rearrangement within Picocoraciae observed when long gaps were excluded unites the “traditional” Coraciiformes. Morphological support for traditional Coraciiformes is mixed; traditional Coraciiformes form a clade in the analyses of Livezey and Zusi [
89], whereas the analyses of Clarke
et al. [
90] conflict. However, we found it provocative that the gap trees support a
Momotus-
Todus clade, a topology that agrees with some morphological analyses [
90,
91] and conflicts with analyses of nucleotide data (
Figure 1). However, none of the analyses of gap data had bootstrap support ≥70% for the
Momotus-
Todus clade.
Figure 5.
(a) Comparison of bootstrap support in trees based on all gap characters and gap characters >1-bp in length. Bipartitions that appeared well supported (≥70% bootstrap) by one analysis and poorly supported (<50% bootstrap) in the other are shaded. Numbers correspond to the following bipartitions: 1. Ardea-Cochlearius-Eudocimus; 2. Alisterus-Psittacula; 3. Chalcopsitta-Platycercus; and 4. Picodynastornites. (b) Comparison of bootstrap support for analyses using all gap characters and RY-coded nucleotide data. The same numbers of informative characters were used in each of these analyses (next page).
Figure 5.
(a) Comparison of bootstrap support in trees based on all gap characters and gap characters >1-bp in length. Bipartitions that appeared well supported (≥70% bootstrap) by one analysis and poorly supported (<50% bootstrap) in the other are shaded. Numbers correspond to the following bipartitions: 1. Ardea-Cochlearius-Eudocimus; 2. Alisterus-Psittacula; 3. Chalcopsitta-Platycercus; and 4. Picodynastornites. (b) Comparison of bootstrap support for analyses using all gap characters and RY-coded nucleotide data. The same numbers of informative characters were used in each of these analyses (next page).
In contrast to the modest differences between analyses using different gap data matrices, much larger differences were observed when we compared bootstrap support from the nucleotide and gap trees (
Figure 5b). This observation does not reflect differences in state space because the nucleotide data were RY-coded to address the more limited character state space in the binary gap characters. These results suggest the existence of both congruent and incongruent signals in the gap and nucleotide data and indicate that the incongruent signals in the gap data were not disproportionately associated with gaps based upon the shortest indels.
3.3. Combined Analyses of Nucleotide Substitutions and Gap Characters
ML analysis of the combined nucleotide and gap character data (including invariant gap characters) resulted in an estimate of phylogeny (
Figure 6) virtually identical to the nucleotide tree (
Figure 1). In general, there was a modest increase in the average bootstrap support for groups in the partitioned ML analyses of nucleotide substitutions and gap characters (
Figure 6). However, there were also five nodes that exhibited more substantial increases in bootstrap support (>10%); four corresponded to supra-ordinal clades (
Figure 6) and the fifth to the
Balaeniceps-
Scopus clade in Pelecaniformes (which increased to 75%). This general increase in support is consistent with the general assumption that including indel information in phylogenetic analyses would prove useful.
Figure 6.
Combined evidence estimate of the avian tree of life. A partitioned ML analysis was conducted using the GTR+I+Γ model for the nucleotide partition and the CFNv+Γ model for the gap partition. Arrows indicate nodes defining supra-ordinal clades where bootstrap support increased or decreased by more than 10% relative to the nucleotide analysis (
Figure 1). The combined evidence topology for Columbiformes was congruent with the gap topology instead of the nucleotide topology (inset; bootstrap values are reported for combined analysis [above branches] and for gap characters [below branches]).
Figure 6.
Combined evidence estimate of the avian tree of life. A partitioned ML analysis was conducted using the GTR+I+Γ model for the nucleotide partition and the CFNv+Γ model for the gap partition. Arrows indicate nodes defining supra-ordinal clades where bootstrap support increased or decreased by more than 10% relative to the nucleotide analysis (
Figure 1). The combined evidence topology for Columbiformes was congruent with the gap topology instead of the nucleotide topology (inset; bootstrap values are reported for combined analysis [above branches] and for gap characters [below branches]).
There were also four nodes in the combined evidence tree that exhibited fairly large (>10%) decreases in bootstrap support. These decreases were evident for three supra-ordinal groups (
Figure 6) and the
Dendrocolaptes-
Scytolopus clade in Passeriformes (which decreased to 66%). There was another difference between the nucleotide and combined evidence trees within Columbiformes. The combined evidence topology for this order corresponded to that in the gap tree, where the relevant branches had even higher bootstrap support (
Figure 6). Although the majority of differences between the nucleotide tree and the gap tree are likely to reflect the more limited power of gap characters to resolve phylogeny, these differences are likely to indicate the existence of conflicting phylogenetic signals in nucleotide substitutions and gap characters. These conflicts are likely to highlight nodes in the Early Bird tree [
13] that should receive additional scrutiny.
There were two nodes with high bootstrap support in both the nucleotide (
Figure 1) and gap trees (
Figure 3) that conflicted; in both cases, the total evidence tree (
Figure 6) was consistent with the nucleotide tree. Surprisingly, given the lower homoplasy of gap characters relative to nucleotide data (
Table 2), independent evidence suggested that the nucleotide tree was more likely in both cases:
The nucleotide tree supports the monophyly of Notopalaeognathae in contrast to both the MP and ML gap trees (
Figure 3), although only the latter had high bootstrap support. The nucleotide topology is strongly supported by independent evidence, including reanalyses of complete mitochondrial genomes [
29], analyses of independent nuclear data matrices [
31], TE insertions [
62] and analyses of morphological data.
The nucleotide tree supports a clade comprising New World quail (
Colinus) and Phasianidae within Galliformes (
Figure 1), whereas the gap tree supports a clade comprising Guineafowl (
Numida) and Phasianidae (
Figure 3B). The former topology is supported by analyses of multiple nuclear and mitochondrial sequences [
8,
83], TE insertions [
92] and morphology [
84].
The combined evidence tree was virtually identical to the nucleotide tree, probably reflecting the ability of rapidly accumulating nucleotide changes to overwhelm the analysis. Nonetheless, the signal in gap characters appears to have an influence, because several supra-ordinal clades exhibited increases or decreases in bootstrap support ≥10% relative to the nucleotide tree. Support for the unnamed clade uniting Novaeratitae and Tinamiformes increased substantially. Although the existence of this clade is supported by analyses of complete mitochondrial genomes [
29] analyses of independent nuclear data [
31] were equivocal and two TE insertions [
62] conflicted with the clade (there were no TE insertions consistent with the combined analysis). Likewise, support for Insolitaves also increased, although there is no independent evidence supporting this clade (
Table 1). Finally, support for Accipitriformes, one the few orders with limited bootstrap support, also increased (from 58% to 71%). In contrast, two supra-ordinal clades with independent corroboration (Australaves and Eucavitaves) exhibited decreased support. Neither of those two clades appeared in the gap tree (
Figure 3). The decreased support for Australaves and Eucavitaves in the combined evidence topology is consistent with the hypothesis that the nucleotide and gap data exhibit some genuine (albeit limited) conflict.
3.4. Analyses of Gap Characters and Models of Indel Evolution
The conflicts between the gap and nucleotide data may reflect the poor fit of the models we used for analysis. Better models of indel evolution are clearly desirable, because the actual patterns of indel evolution are no doubt more complex than the combination of gap coding and analyses using the CFNv+Γ model or parsimony-equivalent models. Indeed, it is unlikely that any of the models used in phylogenetics have a perfect fit to the underlying processes of sequence evolution. Nonetheless, approximating models have proven very useful for phylogenetic estimation (see Sullivan
et al. [
93] and Huelsenbeck
et al. [
81] for additional discussion). Thus, we felt that the simple ML approach we used represented a reasonable starting point that should be tested. However, we did not find this simple ML method performed substantially better than analyses using the MP criterion, suggesting future studies should explore more complex models.
Models of sequence evolution have improved along with our understanding of the processes of sequence evolution [
81]. This raises the question of which aspects of indel evolution might prove to be most important for improving models of indel evolution. Although short indels are more common than long indels [
10,
86], we found that filtering the data matrix to remove short indels did not improve congruence, raising questions about the value of incorporating this correlation into models of indel evolution. The existence of a deletions bias has been established both for birds [
17,
86,
94] and mammals [
11], and incorporating this asymmetry might be useful. Indeed, asymmetry should be intrinsic to models of indel evolution; sequence alignments that represent evolutionary history accurately can include homoplastic deletions, but homoplastic insertions should be forbidden (since distinct insertion are, by definition, not homologous; e.g., Alekseyenko
et al. [
95]). Although more complex and realistic models that combine sequence and indel evolution in this manner have been proposed [
95], it is unclear they can be implemented in a way that will prove to computationally tractable for phylogenies of this size. It also remains unclear whether these more complex models capture all of the relevant features of indel evolution, but the observation that analyses excluding indel information can be positively misleading [
14,
22] suggests that development of improved models of indel evolution remains critical.
Another aspect of model fit that should not be ignored is the assumption that a single tree underlies the observed distribution of gaps. Gene trees can differ from the species tree for several reasons [
96]; for avian phylogeny, the most common reason is probably deep coalescence. The short branches at the base of Neoaves (
Figure 4) suggest that incomplete lineage sorting due to deep coalescence was common during the radiation of this group [
97]. The distribution of TE insertions is consistent with incomplete lineage sorting [
60,
61]. Discordance among gene trees is known to lead to the incorrect estimation of species trees when concatenated analyses are conducted [
98], and we expect concatenated analyses of gaps from multiple loci to inherit all of the properties of similar analyses that use nucleotide data. Although nucleotide and gap data reflect the same genes and, therefore, the same set of gene trees, the number of gap characters and variable nucleotides differs among loci. These differences in the number of characters in each partition effectively result in differential weighting of loci in the gap and nucleotide trees and, therefore, create the potential for analyses of nucleotides and gaps to recover different topologies.